Data Classification

This page explains how the site’s own articles describe data classification. It is not legal advice, and a qualified adviser should confirm what applies to your own data.

What Data Classification is

Data classification is sorting information by how sensitive it is, so that the rule for using AI tools is drawn by the kind of data and not by the name of the tool. The site’s customer data article uses three tiers: green for any tool, amber only for an approved business tool and red kept out of general AI tools.

The idea sounds technical but works as an everyday habit. Before pasting something into an AI tool, ask what the most sensitive thing in it is, and which tier that belongs to. The answer takes about five seconds and, the article says, gives the same result for everyone.

Why draw the line by Data and not by Tool

The usual advice is “never paste customer data”. The customer data article says staff ignore it within a week, because it blocks the tasks that make the tools worth using. The banning article makes the related point that a ban on tools gives no clear edge, because generative AI now sits inside email clients, meeting software, document editors and browsers, often switched on by default.

A rule by kind of data avoids both problems. It tells people what may never be entered and what is fine, and it holds whichever tool appears next. The customer data article puts the useful question in two halves: can this text identify a person or reveal something confidential, and is the tool covered by an agreement the business has made? If the first answer is no, the text can mostly be pasted without worry. If it is yes, the second half decides.

The three tiers

The article sets out the tiers in plain terms.

• Green: safe in any tool, including a personal account. Text that identifies no one and reveals nothing confidential, such as published website copy, public product information, general questions, invented examples and real situations described so generally that no customer could be picked out. The article’s test is whether you would be comfortable seeing the text on your public website.
• Amber: only in a tool the business has approved. Text that contains personal or confidential detail that the task needs, such as a customer’s name and the substance of a complaint, a supplier’s quote or a draft contract with names removed but terms intact.
• Red: stays out of general AI tools. Some categories carry a risk no setting fully removes: login details, passwords, API keys and access tokens, payment card and bank details, identity and passport numbers, health information, information about children, anything under a confidentiality agreement or legal privilege, and complete customer lists. A red task that is genuinely worth automating needs a deliberate design with a data protection assessment, not a quick paste.

The banning article gives a shorter version for staff. It lists what may never be entered into any AI tool, such as client names and contact details, contracts, HR records, credentials and anything under a confidentiality agreement. It lists what is fine in the approved tool, and says anything else goes to one named person first.

The most Sensitive element sets the tier

A text takes the tier of its most sensitive element. One passport number in a three-page document makes the whole document red. The article calls this the step staff find counterintuitive, because they judge by the bulk of the text, and the bulk is often harmless.

Two refinements stop the test becoming a ritual. The first is to ask whether the task needs the personal detail at all. A reply draft needs the complaint, not the name and address, and a summary of a call needs the substance, not the phone number. When the detail is not needed, text can drop from amber to green. The second is that the test applies to attachments and screenshots as much as to typed text. A screenshot carries every field on the screen, and a spreadsheet carries every row and column, hidden ones too.

Why account type matters more than brand

The customer data article says the same assistant behaves differently under a personal login and under a business agreement. It names five things that can happen to a prompt after it is sent, and each maps to a setting or a clause that can be checked.

1. It leaves your systems. The text travels to the provider’s servers, often in another country.
2. It may be stored. A copy of the data sits in an account, and if the account is personal it stays outside the business’s control.
3. It may be used to improve the model. Some consumer products do this unless it is switched off, and business accounts often do the opposite by default.
4. It may be seen by people. Providers may let staff review a sample of conversations, depending on the plan and the contract.
5. It may come back. Memory features, saved instructions and connected folders can bring a detail back in a later conversation.

These are the reasons the four amber conditions exist: the account belongs to the business, a data processing agreement is in place, training on inputs is off or contractually excluded and the business knows how long inputs are kept. Where those are not met, the text is red until they are. The article adds that responsibility stays with the business: it cites the ICO as saying that buying an AI system from a supplier does not remove the business’s own data protection obligations.

Removing Personal Data without losing the task

Often the cheapest safe option is to remove the personal detail before the text goes anywhere. The article gives a method: replace identifiers with placeholders and keep a key outside the tool, remove details that identify by combination, cut what the task does not need, check pasted files and images and re-insert the real details locally afterwards.

The article is honest about the limits. Replacing a name with a code is pseudonymisation, which it says the ICO describes as essentially a security and risk-reduction technique, with pseudonymised data remaining personal data. A record cannot be treated as anonymous because a name was removed. So the article’s advice is to treat a lightly edited record as amber, use an approved tool and keep the key safe. The data protection and GDPR pages cover the duties behind this.

Turning the tiers into a one-page Policy

A rule in one manager’s head protects nobody. The customer data article gives six parts for a one-page rule:

1. Approved tools, with the account type to log into. Everything else is for green text only.
2. The three tiers, with two examples of each from the business’s own work.
3. The placeholder habit, with a before and after.
4. Who to ask, a named person, with a reply time.
5. What to do after a mistake, with reporting early rewarded and not punished.
6. A review date, because tools change their terms often.

The banning article’s three-part version is shorter but the same in spirit. It says one page read does more good than ten that are not, and that the page adds no new principle, because client and HR data is already covered by data protection duties. It makes an existing duty usable on a busy afternoon. Both articles say to explain the reason for each line, because a rule people understand gets followed in cases it did not anticipate. The AI literacy page covers what staff need to judge such a case.

Why a ban does not replace a Data rule

The banning article reports survey figures on what happens where no rule exists. In a July 2026 survey of 500 US adults, 6.6% said their employer bans AI tools outright and 35.8% said their employer has a clear written policy on what work information may be shared. Of the workers surveyed, 38% admitted entering at least one type of work information into a personal AI account their employer does not control. A separate survey of more than 1,000 employees found 28% said they would use AI at work even if it were banned.

The article’s own caution is that both surveys are self-reported and were run by commercial firms, so they show direction and not precise measurement. The direction it draws is that a ban removes the approved route but leaves the task, so the use moves onto accounts the business cannot see. That is shadow AI. The ban on AI tools page covers it.

The article does not say never to ban anything. It says narrow bans are right where a specific tool is unacceptable for a specific kind of data, and that a good one names the tool or the data, gives the reason, points to the approved alternative and says who to ask.

A worked example

An illustration, not a case study. A small business notices that staff paste customer emails into whichever AI assistant they like.

The owner lists the tools in use, including the unapproved ones, and notes the account type behind each. One tool, on a business account, meets the four amber conditions, so it becomes the approved tool. The owner writes a one-page rule with the three tiers and two examples of each from the business’s own inbox. A complaint reply is amber, so it goes into the approved tool with the customer’s name replaced by a placeholder. A request to summarise a public product guide is green and can go anywhere. A scanned passport attached to an onboarding email is red, so the whole email stays out of general AI tools.

The owner names a person to ask, promises an answer within the hour and says that reporting a wrong paste early will not be punished. A short session follows, where the team applies the tiers to real examples from their own inbox and argues about the borderline cases. The pasting has not stopped, but it now runs through a tool the business can see.

Mix-ups worth avoiding

Four confusions come up often enough to name. The first is judging a text by its bulk when the most sensitive element sets the tier. The second is treating a removed name as anonymisation, when the article says pseudonymised data remains personal data. The third is assuming the vendor carries the responsibility, when the article says buying a tool does not remove the business’s own duties. The fourth is treating a ban as the data rule, when a ban removes the approved route and leaves the task.

How we use it

The customer data article suggests teaching the three tiers in a short session where the team applies them to real examples from their own inbox, because the borderline cases are where understanding forms. Following each team’s data through its processes is part of how we audit.

If you would like an outside view on which tools your team uses and what goes into them, a discovery call is a quick way to start. An audit is a diagnostic and not a legal opinion.

Sources

• Customer data and AI tools: what is safe to paste: the three tiers, the four amber conditions, the most sensitive element, the five things that happen to a prompt, the placeholder method and its limits, and the one-page rule.
• Why banning AI tools at work does not work: drawing the line by data, the one-page rule, why a ban moves use out of sight, narrow bans and the survey figures with their caveat.
• Ban on AI tools, Data protection, GDPR, Shadow AI, AI governance, AI tool inventory, AI literacy, Data residency, Human in the loop and Risk classification: the neighbouring terms this page links to.